Are Multiple Organs, Eyes, Teeth, Lesions, or Events from One Participant Independent Observations?

If One Participant Contributes Several Organs, Lesions, or Events, Can You Count Them as Independent Observations?

Suppose an ophthalmology study recruits 100 patients and measures both eyes, producing 200 eye observations. A dental study examines 500 teeth from 80 patients. Another study records 300 lesions from 120 patients or several hospitalizations for each person.

These datasets contain more observations than participants. Each eye, tooth, lesion, or event may have distinct characteristics and outcomes. But observations belonging to the same person may also share biological, behavioral, environmental, and treatment-related influences.

Can researchers analyze every observation as an independent case? Or does the shared participant create a dependence structure that must be considered when estimating effects, standard errors, and sample size?

Distinct Measurements From One Participant Are Not Automatically Independent

How to Analyze Multiple Measurements or Events From the Same Participant

Why Are Observations From the Same Person Often Related?

Measurements from the same participant may be correlated because they share characteristics that influence several observations simultaneously.

Two eyes belong to the same person and may share genetic, metabolic, environmental, and systemic disease influences. Multiple teeth may share oral hygiene practices, dietary exposures, socioeconomic conditions, and treatment histories.

Likewise, several lesions within one patient may be influenced by the same immune system, underlying disease process, or medication exposure.

Repeated clinical events may be related because an individual's underlying susceptibility persists over time or because earlier events affect the likelihood of later events.

These relationships mean that observations within a participant may provide less independent information than an equal number of observations from unrelated participants.

Does Correlation Mean the Observations Are Identical?

No. Correlation and identity are different concepts.

A person's left and right eyes may have different visual acuity, disease severity, or treatment responses. Teeth within the same mouth may differ in condition, anatomical location, or exposure history.

These differences can be scientifically important.

The presence of within-participant correlation does not mean researchers must discard all but one observation. It means that the statistical analysis should represent the dependence appropriately.

For example, both eyes may be included in an ophthalmological analysis using a method that accounts for clustering within patients.

The appropriate strategy depends on the outcome, the level of exposure, and the estimand.

What Is the Unit of Analysis in These Studies?

The answer depends on the research question.

If the study concerns patient-level outcomes, such as whether a person experiences any complication, the participant may be the principal unit of analysis.

If the question concerns tooth-level treatment outcomes, individual teeth may be the observational and outcome units, with teeth clustered within patients.

If the question concerns lesion-specific response to treatment, lesions may be the outcome units while participant-level clustering remains relevant.

Researchers should distinguish the unit of analysis from the number of observations and the level at which statistical independence is assumed.

Multiple observations from one participant do not automatically create multiple independent participants or experimental units.

How Do Different Clinical Data Structures Compare?

Some studies involve several dependence levels simultaneously. For example, repeated measurements of multiple teeth within each patient produce measurements nested within teeth and teeth nested within patients.

A single participant identifier may not fully describe this structure. Researchers may also need tooth, lesion, organ, or event identifiers.

What Is Within-Participant Correlation?

Within-participant correlation describes statistical dependence between observations belonging to the same person.

For continuous outcomes, an intraclass correlation coefficient may summarize the extent to which observations resemble one another within participants under a specified model.

However, the appropriate correlation measure depends on the outcome and structure. Correlation between two eyes is not necessarily represented by the same model as correlation among repeated hospitalizations.

For bilateral measurements, left and right observations may also differ systematically because of anatomical or clinical characteristics.

Researchers should examine whether the dependence structure is approximately exchangeable, varies by anatomical location, or changes with time.

Statistical independence should not be inferred simply because two observations have different values.

What Happens If You Ignore Within-Participant Dependence?

Ignoring correlation can produce incorrect standard errors, confidence intervals, and hypothesis tests.

When observations are positively correlated, treating them as independent commonly overstates the effective information available for estimating a pooled mean or certain exposure effects.

For example, 200 eyes from 100 patients may provide less independent information than 200 eyes from 200 unrelated patients for some estimands.

The exact consequences depend on the correlation, group sizes, outcome distribution, and analytical model.

Incorrect independence assumptions may also affect the estimated association itself when the analysis fails to distinguish patient-level and organ-level predictors or gives disproportionate weight to participants contributing more observations.

These concerns are related to misidentifying the unit of analysis and treating correlated observations as independent evidence.

Can Both Eyes Be Included in an Ophthalmology Study?

Yes. Both eyes may contribute valuable information, and excluding one eye is not universally necessary.

However, researchers should account for inter-eye correlation when the outcome and analytical method require it.

Several strategies are possible. Researchers may select one eye per participant using a prespecified rule, summarize outcomes at the patient level, or include both eyes in a correlated-data model.

Selecting one eye can simplify analysis but may discard information and reduce precision. Selecting the more severely affected eye can introduce selection concerns or change the target population and estimand.

Including both eyes using generalized estimating equations or mixed-effects models may preserve information while accounting for dependence.

The preferred strategy depends on whether the research question concerns eyes, patients, bilateral disease, or treatment effects.

Can Multiple Teeth From One Patient Be Analyzed Separately?

Yes, when the outcome concerns individual teeth and the analytical method appropriately represents their dependence.

For example, a dental study may investigate whether tooth-level characteristics predict restoration failure.

Each tooth can contribute distinct information about anatomical location, restoration type, or local disease conditions.

However, teeth within the same patient share behaviors, biological characteristics, and healthcare access.

Ignoring this clustering may produce misleading uncertainty estimates.

Some dental datasets also contain repeated examinations of the same teeth, creating a hierarchy of observations within teeth and teeth within patients.

Multilevel models or other methods appropriate to the outcome and dependence structure may be required.

What About Multiple Lesions Within One Patient?

Lesion-level analysis may be appropriate when the scientific question concerns lesion characteristics or lesion-specific treatment response.

However, multiple lesions from the same patient may share disease biology, immune responses, treatment exposure, and other patient-level factors.

For example, if every lesion in a patient receives the same systemic treatment, the lesions do not constitute independently assigned treatment units.

Analyzing them as though each lesion independently received the intervention may create pseudoreplication.

If lesions receive different locally assigned treatments within the same participant, the assignment structure differs and may support a within-patient comparison, subject to interference and other design considerations.

Researchers should distinguish lesion-level outcomes from patient-level treatment assignment.

Are Repeated Clinical Events the Same as Multiple Organs?

Not exactly. Both involve observations clustered within participants, but recurrent events introduce additional temporal and event-process considerations.

For example, a patient may experience several hospitalizations, disease relapses, or falls during follow-up.

These events occur at different times, and earlier events may influence subsequent risk.

The relevant estimand might concern time to first event, total event burden, recurrent-event intensity, or the probability of experiencing a specified number of events.

Different questions require different analyses.

Methods for recurrent events may include Andersen–Gill models, event-order-specific approaches, frailty models, or suitable count models, depending on the event process and assumptions.

Simply treating each event as an independent participant-level outcome ignores both clustering and potentially important temporal relationships.

What If One Participant Contributes More Observations Than Another?

Unequal numbers of organs, lesions, teeth, or events can affect the interpretation of an analysis.

Suppose one patient contributes ten lesions while another contributes one. An unweighted lesion-level analysis gives the first patient ten times as many lesion observations.

This may be appropriate if the estimand concerns the average lesion in the observed lesion population.

It may be inappropriate if the estimand concerns the average patient, because patients with more lesions contribute more strongly.

Researchers should distinguish participant-weighted and observation-weighted estimands.

When cluster size is associated with outcomes, a phenomenon often called informative cluster size, standard approaches may require modification to estimate the intended population quantity.

Appropriate weighting or specialized models may be considered, but their assumptions should be made explicit.

Can You Randomly Select One Observation per Participant?

Yes. Selecting one eye, tooth, lesion, or event per participant can eliminate certain forms of within-participant clustering from the resulting dataset.

However, this approach discards potentially useful information and may change the estimand.

Selection should follow a prespecified and scientifically defensible rule.

For example, randomly selecting one eligible eye may be preferable to selecting whichever eye has the most favorable outcome.

Choosing the first event rather than all recurrent events changes the research question from event burden to first-event occurrence.

Researchers should not select observations after examining which choice produces the most favorable statistical result.

Can You Average Multiple Observations Within Each Participant?

Participant-level aggregation may be appropriate when the research question concerns a meaningful summary of an individual's condition.

For example, a study may examine mean visual acuity across both eyes or average lesion size per patient.

However, an average can conceal clinically important differences between organs or lesions.

It may also be inappropriate when the number of observations varies substantially or when outcomes are nonlinear, categorical, or time-dependent.

Averaging several hospitalizations into one value is not equivalent to modeling recurrent-event risk.

Researchers should justify the summary measure and establish that it corresponds to the intended patient-level estimand.

When Are Generalized Estimating Equations Appropriate?

Generalized estimating equations, or GEE, can estimate marginal associations while accounting for correlation among observations within participants.

They may be useful for binary, continuous, or count outcomes when multiple organs or anatomical sites contribute observations.

For example, a GEE logistic model could examine a binary eye-level outcome while treating patient identity as the clustering variable.

The working correlation structure should be appropriate to the data. Robust variance estimation can provide protection against certain correlation misspecifications under suitable large-sample conditions.

However, conventional sandwich standard errors may be unreliable with few independent participants or clusters.

GEE does not automatically address confounding, informative cluster size, or incorrectly defined outcomes.

When Are Mixed-Effects Models Appropriate?

Mixed-effects models can represent participant-specific heterogeneity through random effects and may accommodate multiple nested levels.

For example, a study measuring several lesions repeatedly may use a model with observations nested within lesions and lesions nested within patients.

A random patient intercept can represent shared patient-level influences, while additional random effects or covariance structures may represent lesion-level dependence.

Mixed-effects models can be useful when researchers want to estimate conditional associations or investigate variation between participants.

However, the interpretation of conditional effects may differ from marginal effects, especially for nonlinear models.

Model selection should therefore reflect the intended estimand rather than the assumption that one method is universally superior.

What Is the Difference Between GEE and Mixed-Effects Models?

Both approaches can be appropriate, but they do not necessarily estimate the same parameter.

Researchers should specify whether their scientific interest concerns an average observation in the population, a participant-specific relationship, or another clearly defined quantity.

What If Treatment Is Assigned at the Patient Level?

When treatment is assigned to patients, all eligible organs or lesions within a patient may receive the same treatment condition.

In such studies, the patient is the unit of treatment assignment, even if outcomes are measured for multiple organs or lesions.

For example, if 40 patients are randomized to two systemic treatments and each contributes several lesions, the study has 40 independently randomized patients.

The lesions provide outcome information but do not constitute additional independent treatment assignments.

Failure to distinguish these levels can create pseudoreplication and misleading treatment-effect inference.

The distinction between randomization and measurement units should be established during study design.

What If Treatment Is Assigned Separately to Different Organs?

Some studies assign treatments within participants. For example, one eye may receive one treatment while the fellow eye receives another.

This creates a paired or within-participant experimental structure.

Such designs may control for stable patient-level characteristics and improve precision for certain treatment comparisons.

However, the analysis must preserve the pairing, and investigators should consider whether treatment effects can spill over between organs.

Systemic absorption, shared behavioral responses, or other interference mechanisms may complicate the assumption that one organ's treatment does not affect another organ's outcome.

The appropriate experimental unit and causal estimand depend on how treatment is assigned and whether interference is plausible.

How Should Sample Size Be Calculated?

Sample size planning should distinguish the number of participants from the number of organs, lesions, or events contributed by each participant.

Precision depends on the intended estimand, within-participant correlation, variation in observation counts, and treatment assignment structure.

For example, recruiting more patients may provide additional independent biological variation, while measuring more lesions within existing patients may improve precision for some lesion-level outcomes.

The relative benefit depends on the correlation and the research question.

When treatment is randomized at the patient level, the number of randomized patients is particularly important for treatment-effect inference.

Simulation-based sample size planning may be useful when the design includes unequal cluster sizes, multiple dependence levels, or recurrent events.

How Should Multiple Observations Be Reported?

Report the number of participants and the total number of organs, teeth, lesions, or events separately.

Describe how many observations each participant contributed and whether those counts varied substantially.

Methods sections should identify the observational unit, outcome unit, treatment assignment unit, and dependence structure.

Explain whether observations were selected, aggregated, or modeled using a correlated-data approach.

When applicable, report assumptions about within-participant correlation, informative cluster size, missing observations, and repeated events.

These details help readers assess whether the analysis avoids overstating the independent information available from clustered observations.

Analyzing Both Eyes in a Study of Visual Outcomes

For illustration, suppose the intraclass correlation between fellow-eye outcomes is 0.50. Under a simple exchangeable-correlation approximation with two eyes per patient, the design effect is:

The calculation illustrates why 200 correlated eye measurements should not automatically be treated as equivalent to 200 independent eyes from unrelated patients.

The actual precision of the anatomical-characteristic association would depend on its distribution within and between patients, the outcome variance, correlation structure, and selected model.

If the anatomical characteristic varies mainly between patients rather than between fellow eyes, the available information for estimating its association may differ substantially from a predictor that varies independently between eyes.

Common Mistakes With Multiple Organs, Lesions, and Events

How to Choose the Appropriate Analysis for Multiple Observations per Participant

Start by identifying the scientific entity of interest. Determine whether the question concerns patients, organs, lesions, or recurrent events.

Then examine the number of observations per participant, their dependence, and the level at which exposures or treatments are assigned.

Multiple observations from one person can strengthen research when their structure is used appropriately. The objective is not to discard useful measurements but to avoid presenting correlated information as independent replication.

What to Verify Before Analyzing Multiple Observations From One Participant

Frequently Asked Questions About Multiple Observations From One Participant

Distinct Anatomical Observations Do Not Necessarily Provide Independent Evidence

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.